TRANSACTIONS of the AMERICAN MATHEMATICAL SOCIETY Volume 240, lune 1978
ERICSSON'SCONJECTURE ON THE RATEOF ESCAPEOF ¿-DIMENSIONALRANDOM WALK BY HARRYKESTEN
Abstract. We prove a strengthened form of a conjecture of Erickson to the effect that any genuinely ¿-dimensional random walk S„, d > 3, goes to infinity at least as fast as a simple random walk or Brownian motion in dimension d. More precisely, if 5* is a simple random walk and B, a Brownian motion in dimension d, and \j/i [l,oo)-»(0,oo) a function for which /"'/^(OiO, then \¡/(n)~'\S*\-» oo w.p.l, or equivalent^, ifr(t)~'\Bt\ -»oo w.p.l, iff ffif/(tY~2t~d/2 < oo; if this is the case, then also $(ri)~'\S„\ -» oo w.p.l for any random walk Sn of dimension d.
1. Introduction. Let XX,X2,... be independent identically distributed ran- dom ¿-dimensional vectors and S,, = 2"= XX¡.Assume throughout that d > 3 and that the distribution function F of Xx satisfies supp(P) is not contained in any hyperplane. (1.1) The celebrated Chung-Fuchs recurrence criterion [1, Theorem 6] implies that1 |5n|->oo w.p.l irrespective of F. In [4] Erickson made the much stronger conjecture that there should even exist a uniform escape rate for S„, viz. that «""iSj -* oo w.p.l for all a < \ and all F satisfying (1.1). Erickson proved his conjecture in special cases and proved in all cases that «-a|5„| -» oo w.p.l for a < 1/2 — l/d. Our principal result is the following theorem which contains Erickson's conjecture and which makes precise the intuitive idea that a simple random walk S* (which corresponds to the distribution F* which puts mass (2d)~x at each of the points (0, 0,..., ± 1, 0,..., 0)) goes to infinity slower than any other random walk. Theorem. Let d > 3, S„ and S* as above, and let [B,)t>0 be a d-dimen- sional Brownian motion (EB, = 0, EB,(i)Bt(j) = t8¡f). Assume that the func- tion \[i: [ l,oo) -» (0,oo) satisfies rl/\(t)iO. (1.2) Then \¡/(n)~x\S*\-» oo w.p.l and\p(t)~x\B,\ -> oo w.p.l are both equivalent to
Receivedby the editors June 25, 1976. AMS (MOS) subjectclassifications (1970). Primary 60J15, 60F20. Key wordsand phrases. Random walk, escape rate, concentration functions. 'For any vector c e R2(/)}l/2."w.p.l" stands for "with probability one". O American Mathematical Society 1978
65
License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 66 HARRY KESTEN f^(t)d-2rd/2< oo. (1-3)
If (1.2) and (1.3) hold, then also ^(/z)-1!^! -» oo w.p.l whenever F satisfies (1.1). As we shall see this theorem is almost immediate from Proposition 1. If d > 3 and F satisfies (1.1) then there exist constants 0 < r,(P), T2(P) < oo such that for all k > 1 and all A > 0, P {|Sn| < A for some 2k < n < 2k+x)
< TX(F){((A + l)2-k/2)"-2 + exp -T2(F)k}. (1.4) It is worthwhile contrasting the theorem with the observation (see [4, end of §5]2)that even if (1.1)—(1.3)hold, there may exist a deterministic sequence of vectors an such that hminf^(n)-x\Sn-an\=0 w.p.l. (1.5) Thus, the theorem is not merely a matter of estimating concentration func- tions for S„. Nevertheless the result is intimately connected with concentra- tion functions. Erickson's proof for \p(t) = t", a < 1/2 - \/d, uses a con- centration function inequality of Esseen [6], and our proof makes heavy use of the following inequalities for concentration functions in dimension two. These estimates too are closely related to those of Esseen [5, Theorem 3], but are more generally applicable. We note that Corollary 1 shows that in the identically distributed case the concentration function decreases like n~\ Proposition 2. Let Z„ Z2,... ,Z„ be independent random two-vectors with distribution functions Gx, G2,..., G„. Let p, p„ p2,..., p„ be strictly positive numbers such that p, < p, and let AX,A2,... ,A„ be sets in R2 X R2 such that A¡ C {« G R2: |w| > p,}2.Define the symmetrizationGf of G¡ by* G*(B)=fGi(B + u)dGi(u) (1.6) and pur
q¡ = ¡ dGf(u), a2= inf / (0,u)2dG*(u), J\u\>o, 1*1-1•>!<&
C=ï{o2 + £[ dG°(u)dG?(v)\. (1.7) ,-1 I ?, Ju,t>eA, J
2More details are given in the recent article by K. B. Erickson, Slowing down d-dimensional randomwalks, Ann. Probability5 (1977),645-651. 3For a distribution function F and Borel set A, F (A) denotes the mass assigned to A by the Borel measure induced by F. 4<«,o> denotes the usual inner product of two vectors of the same dimension.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use ¿-DIMENSIONALRANDOM WALK 67
Finally, denote the angle between two vectors u and v, determined in such a way that it lies in [0,tt], by
sup P 2z,+
Kxp2 < exp (¿¿-i dGl(u)dG¡(v) \¡=\ Lq¡ Ju.vSA,
•log 1 + (1.8) \ \v\ \sin
Corollary 1. Let Z„ ... ,Z„ be independent random two-vectors, all with the same distribution function G, and define Gs as in (1.6). If p > 0 and q = f dGs (u), a2 = inf f <0,«>2dGs (u),
and if there exist sets BX,B2c [z E R2: \z\ > p) and a constant c > 0 such that \v\ ¡sin
c whenever u E Bx,v E B2, (1.9) then
sup P 2z,+
<— p2(l+pc-x)L2 + £ ( dGs(u) [ dGs(v)\ .(1.10) n [ q JBi jb2 )
Acknowledgement. We wish to thank Professor Erickson for making his results [4] available to us before publication and for some helpful conversa- tions. 2. Reduction to a two dimensional problem. In this section we shall reduce the proof of Proposition 1 to a two dimensional problem. This problem will then be settled together with Proposition 2, by means of estimates on two dimensional concentration functions in §3. Throughout Kx, K2,... will be strictly positive constants which depend on the dimension d only and whose numerical value is immaterial for our purposes. They will not necessarily have the same value at each appearance. r„ T2,... and kx,1> ^2>k will be con- stants which depend on F and d or some other parameters; where necessary these parameters will be indicated explicitly, except that we shall not indicate dependence on d. For any random variable Y we write Ys = Y — Y' where
License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 68 HARRYKESTEN
Y' is independent of Y and has the same distribution as Y. If Y has distribution function G, then Ys has the distribution function Gs given by G'(B)=fG(B+y)dG(y). The first reduction shows that we may assume that F has certain smooth- ness properties. It is of a purely technical nature and has little to do with the basic idea of the proof of Proposition 1. The reader should skip the proof of Lemma 1 at first reading. Lemma 1. It suffices to prove Proposition 1 in the case where F(dx) > adx on C0 (2.1) for some a > 0 and some closed cube C0 C Rd (which does not reduce to a point). Proof. Let F satisfy (1.1). We shall construct an P, which satisfies (2.1) (and a fortiori (1.1)) such that Proposition 1 holds for the original S„ as soon as it holds for a random walk, S„(FX)say, whose increments have distribution function P,. For this purpose we first find a large cube C, = [z E Rd:-L < z(i) < L,l< i < d) (2.2) such that F(CX) > 0 and such that the intersection of supp(P) with C, is not contained in any hyperplane. This is possible by (1.1). Define the distribution functions G and H by5 Po = ÏF(Ci), G (S) = (2p0)-xF(S n C,) and ,,, , H(S) = (i -/><,)-'{ÎW n c,) + F(s n C{)}. Then we can write F = p0G + (1 — p0)H, and if IX,I2,..., UX,U2,..., VX,V2,... are totally independent random variables with P{lj = 0) = l-P{lj=l}=Po, P{UjES} = G(S), P{VjES) = H(S), then the joint distribution of {X„)„>x is the same as of {U„(l — I„) + V„I„)„>X(compare [9, p. 1184]).We shall therefore assume that U,V and / are defined on our original probability space and that X„ = Un(\ — I„) + VnIn. Let o, < a2 < ... be the successive (random) indices for which /„ = 1. Then, with o0 = 0, the oi+x — o, are independent and all with the distribution P{oi+x - a,. = r) =Por"'(l -Pol r > 1. (2.3)
5For any set S c R^, Sc denotes its complement, i.e., Rrf\ S.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use ¿/-DIMENSIONALRANDOM WALK 69
Even more, if % is the o-field generated by { Uj, V},I}: 1 < j < n), then for each {%) stopping time Tand o*(T) = smallest o¡ which exceeds T, one has P{o*(T) = T+ r\5T) =Por-,0 "Po). r > 1. It is also not hard to see that the {Sc+i — So}i>0 are independent and all with the same distribution function6 F2= 2 P{°x = r)G*-» * H= 1 ^-1(1 -Po)G^-~) * H, r-1 r-1 and that also, conditional on fr, Sa,fT)- ST has the distribution F2 on the set {T< oo}. Now take T = ini{n E[2k,2k+X) mth\Sn\< A) (= oo if no such T exists). Then, since |5^r| < A,
P{\S0¡\< 2A for some/ e[I(l -/>0)2*,2(1 - Po)2k+2)}
> P{T
Now for some yl0 = ^r/^") < °° an^ a31A > A&
F2({z ERd: \z\\, P{T
/.=[i(l-/>o)2*], /2-[2(l-Po^+2]. the last term in (2.4) is at most
P{a,t > 2* or a,2 < 2k+2} < T3 exp -T42*,
for some r3,r4 < oo depending onp0 only (by (2.3) and standard exponential
6G * H denotes the convolution of G and H, and C*(I) denotes the i-fold convolution of G with itself.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 70 HARRYKESTEN
estimates; compare [9, formulae (5.40) - (5.42)]). Thus (2.4) yields for A > Aq, P{\S„\ < A for some 2* < n < 2*+1} = P { T < oo} < 2P {\S„\ < 2A for some / G [ \ (1 - />„)2*,2(1- p0)2k+2)} +2plk + 2T3exp - T42*. (2.5)
It is clear from (2.5) that if Proposition 1 holds for the random walk S„(Fj) = S„, whose increments SL — Sa¡ all have the distribution F2, then Proposition 1 also holds for the original S„. Actually (2.5) was derived only for A > A0. However, for A < A0 or even A < 2k/s, (1.4) is immediate from Esseen's estimate [6, Corollary to Theorem 6.2]
P{\S„\ < T5(F)(A + \)dn-á'\ (2.6) for some T5(F) < oo. F2 itself does not have to satisfy (2.1). However, set G2-%Prl(l-Po)G«'-l\ (2-7) r=l so that F2 = G2* H, and assume that we can find a distribution function 07, with a continuous density such that f x" dGx(x) = [ x" dG2(x) for ||i>||< 16. (2.8) (Here we use the standard multi-index notation; x" = Ud.xx(i)Hi),for posi- tive integers v(i) and ||i-|| = 2f,,K0-) ^e c^m triat x^ienF\ = Gx * H has the required properties. Before constructing Gx we shall prove this claim. (2.1) is obvious for P,; indeed Gx has a continuous density and hence P, has a density which is lower semicontinuous. Thus we merely have to prove that the validity of Proposition 1 for Sn(Fx) implies the validity of Proposition 1 for S„(F2) (since we already showed that Proposition 1 then also holds for our original S„). Now fix k and A > 2k/s; we already saw above that (1.4) follows from (2.6) if A < 2k/i so that these are the only values of interest. Let #, Xu 30»• • ■ De independent random ¿/-vectors, also independent of {Sai)l>0, and such that N has a normal distribution with mean zero and covariance matrix k~xA2times the identity matrix, and such that each x¡ has distribution P,. Then, for any n, License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use ¿/-DIMENSIONALRANDOM WALK 71 P{|S0J< 2Á] < P{|50ii0)|< 2^,1 < j < d] < P{\San(j)+ iV0)|< 3^,1 A2 < P{\S„m(j)+ N(f)\<3A,\ Similarly 2 *(/)+ *(/)< 3,4,1< f < d 1-1 < 4dx'2A + K3 exp - ^ < p 2xi (2.10) If 'r'are tne characteristic functions of, respectively, GX,G2and H, then the characteristic functions of Sa + N and 2,Xi + N are exp(-£|0|2) Thus, by the inversion formula, p{|S^C/)+ #C/)|<3¿,l - P <3A,l < j < d /-i2xC/) + *C/) ¿ I sin 3(L4 0,• • •#„ n ' °j J-'\ •| < ( t )vr • • • /_>'( But, by virtue of (2.8), License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 72 HARRY RESTEN M*) - 45 r(e,xy ,«0,x> d(Gx(x) - G2(x)) r=0 ±f<0,x)X6d(Gx(x) + G2(x))= -^f<0,x}X6dG2(x) <-ih\9\2 1/1,16 ^Po-l^-Po)((r-W 16 16! r-i (see (2.7) and recall that G is concentrated on C,). Consequently, for A > 2*/8 and n < 2k+2, P{\S0ii(j) + N(j)\<3A,Kj -P 2 xU) + ^CO < 3¿, i< / < ¿ i-l r6(P)^"/ • • • J>|,6exp(- £ |0|j ¿0,. • • 4, < T1(F)nAd-d-X6k^d+X6^2< r7(P)^C+16)/22-fc+2. Combinedwith (2.9)and (2.10)this yields < 4dx'2A + T¿F)e-k'\ (2.11) P{\S0j<2A} Interchanging the subscripts 1 and 2 we prove similarly P{\SajP Sx*<¿¿"1/2>l -r8(P)e-*/2, (2.12) for all A > 2*/8, n < 2k+2.Finally, to prove our claim we appeal to the proof of Theorem3 in [9].Just as in the estimateson pp. 1179-1181of [9], P{\Sa} < A for some 2k < n < 2*+1} < 2P{ftJ<¿ 2 p{\Sa}<2A), (2.13) \n = 0 )'2*+2_,n-2" and also License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use ¿/-DIMENSIONAL RANDOM WALK 73 2x < 4dx'2A for some 2k < n < 2*+2 2* + 2 -' 2*+2-l2 p 2^ 2x < %dx'2Aw 2x, < 4dx'2A n=0 1 J « = 2* 2* 12*+2_, 2 p > K4 2^ 2x <{d~x/2A 2x <4dx'2A . (2.14) n-0 i n-2k i Since 2- 2 P{|Sj<,4}->oo as/c->ooand,4 >2*/8, n=0 it followsfrom (2.11)-(2.14)that for /c > rc,(P), P {|50J < A for some 2* < » < 2*+1} -i < 4dx/2A for some 2* < n < 2k+x <2K4 2xi + P < 4dx'2A for some 2*+1 < « < 2*+2 2x,i + r9(F)exp(-rc/2). Since 2"x, can be taken for S„(FX)(its increments Xnhave distribution Fx), we see from this that if Proposition 1 holds for Sn(Fx), then it also holds for S„(F2) and for the original S„. (Note that we can always obtain (1.4) for k < kx by increasing T,.) To complete the proof of the lemma we show finally that there exists a Gx with a continuous density and satisfying (2.8). Let there be M positive integer ¿/-vectors v with 1 < \\p\\ < 16 and consider the set 9fc = {(*,). distributionfunction P on Rrfand all 1 <||i»||< 16J, where the constants A„ < oo are chosen large enough such that p^^fx'dG^x) G[-AA], 1 License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 74 HARRY RESTEN Assume first that p E 3 9H. Then there exists a supporting hyperplane of 911 at p (see [3, Theorem 8]), i.e., there exist constants c0,c„, 1 < ||i>|| < 16, not all zero, such that 2 cvzv+ c0 > 0 for all z G <3H, (2.15) 1<|M|<16 and 2 cfp„ + c0 = 0. (2.16) l<||,||<16 Then, if M P(x) = 2 c„*'+ c0. ,= 1 (2.15) implies that fP(x)dR(x) > 0 for all distribution functions R on Rrf with 1/x"dR(x) 1. P(0) = 0 shows that c0 = 0. Moreover, if y„ ... ,yd+x are (d + 1) points in supp(G) = supp(P) n C, which do not lie in any hyperplane, then also, for any integers kx > 0,..., kd+x > 0, 2&¿y, G supp(C72)and hence S cp(kx + •■■ + kd+xyX6(kxyx + •■■+ kd+xyd+xy= 0. 1<|H|<16 If we let k¡ -» oo such that d+i k¡(kx+ • • • + kd+x) '-*\ > 0 with 2 \= h we obtain 2 cy(Xxyx+ ■•■ +Xd+xyd+xy=0 H-16 for all A,> 0 with2fí,'A, = 1,i.e., 2 c„x"= 0 H =16 for all x in the closed convex hull of yx,... ,yd+x. Since these (d + 1) points do not lie in any hyperplane, their convex hull has a nonempty interior (see [3, Theorem 4]) and we conclude c„ = 0 for \\v\\ «■ 16. But then also License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use ¿/-DIMENSIONALRANDOM WALK 75 2 cf(kx + ■■■ + kd+xyl5(kxyx + ■■■ + kd+xyd+x)'= 0, 1 License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 76 HARRY KESTEN then t(u,k) P{\(Sn,o>)\<2Ak} > P{\(Sn_x,«y\ whenever k > k2(F), n < T(k) and w = wy- with 2 < / < d (2.25) for a suitable k2(F) < oo. Lemma 2. There exists a Txo= TX0(F)< oo such that for all |w| = \,k> 1, *(«,*) < TX0A2. (2.26) Moreover there exist a universal K0 < oo an*/ a k3 = k3(F) < oo íwcA that for all \u\ = l,k > k3, T(k)P{\(XX, 'For a random variable Y, o2{ Y) denotes its variance. /[ ] denotes the indicator function of the event between square brackets. License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use ¿/-DIMENSIONALRANDOM WALK 77 Proof. For any real random variable Y its concentration function Q(Y,') is defined by ß(y,L)=supP{y < Y< y + L}. y It is easy to see and well known (see [7, Chapters 1.1 and 2.1]) that this sup is taken on for somey, that for 0 < L, < Z^, Q(Y,L2) < (LXXL2+ I)Q(Y,LX), (2.31) and that for Yxand Y2independent, ß(T, + Y2,L)< min{Q(Yx,L),Q(Y2,L)). (2.32) If Yx, Y2,... are independent and identically distributed with distribution function G, then Theorem 3.1 of Esseen [6] gives ß|2 Y{,L] T2, = L-2 inf f This implies (2.26). The same argument shows for any L > L, P{\(Sn,tS)\ License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 78 HARRY RESTEN P{|<5r_„(o>| < M(a)Ak) >l-(Sd + 8)"1 >\ (2.34) (see (2.21) for w ^ w2, and the definition of T(k) and M(u2,k) = 1 for a = u2). On the other hand, by (2.31) with L, = ¿ M(u>)Akand (2.33),the left-hand side of (2.34)is bounded above by i7^(r-ir,/2p{|<^,tó>|>¿M(wK}-1/2, so that (T- \)P{\(Xi¿>\>\M(0)Ak} < 2%2. (2.35) Now (2.32) for Yx= Y,Y2= T shows that ß(F*,L) < Q(Y,L) and this, together with (2.35),yields ß«*„ P{\XX\>TX2}<\, (2.37) and k3 = k3(F) so large that for all k > k3, \AkMM(<*,k) > r12 as well as (T(k) - iyx2x2KJ 2. Then (2.36) shows that there exists some interval [a,a + jM(co)Ak] which contains P{| This proves (2.27)for k > k3(F). To obtain (2.28) we again apply (2.33) with Y¡= <*„w> and take L = M(tS)Ak.Combined with (2.31)and (2.34)this gives 3KS(T- \yx/2M(u)Ak{( License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use ¿/-DIMENSIONALRANDOM WALK 79 T(k) Í d-[( dG(y)}l( y¿/C7(y) = £{y||T| Then \d\ < c and ( y2dG<(y) > // (y, - v2)2 - // {(J, - d)2 + 2(yx - d)(y2 - d) + (y2 - d)2} dG(yx) dG(y2) \yt\ = 2f dG(y2)f (yx-d)2dG(yx). Consequently, a2{y/[|y| = f (y-d)2dG(y) + d2[ dG(y) J\y\ <ï(P{\Y\< c})~-[ y2dG>(y) + P{|y|> c) c2. J\y\<2c When this is applied to Y = (Xx,a), c = M(oi)Ak, we obtain from the above inequality: T(k)o2{(Xx,u)l[\(Xx,«>)\< M(u>)Ak]} <(26K2(P{\(Xx,œ)\ + T(k)P{|<*„W>|> M(o>)Ak}){M(u)Ak}2. Togetherwith (2.27)and(2.29) this implies(2.28). Lastly we prove (2.30). For ¿o = ¿o2(2.30) is immediate since we already observed that (2.22) holds for ¿o= ¿o2with N replaced by T. We therefore fix u =£ u2 for the remainder of the proof. Since, for any n < T, (ST,u>}is the sum of the independent random variables <5[7>I-i]n,to>and License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 80 HARRY RESTEN P{\(ST,cc}\<^2M(u>)Ak} <33Q{{ST,u),3\M(u)Ak) (by (2.31)) < 33Q{(S[Tn-,xn, < 33K5[T/n]-l'\p{\(Stà\>±M(<*)Ak})-l/2 (by(2.33)) < 33K5[T/n]-}/2{l - Q((Sn,Uy,^M(o:)Ak)y~/2 (as in (2.36)). (2.38) Now in order to prove (2.30) we only have to consider those ¿ofor which P{|<5r,|>ÍM(«K}<¿. For such an ¿o(2.38) implies for all n < T, Q{(S„,U},¡M(w)Ak) > 1 - 2X4K2nT~x. (2.39) Now apply (2.39)with n = N', where N' = N'(u,k) = min(«: P{| > p{(Su+iw - SjN.,w}>\M(a)Ak, 0 > (l - 2%2^)r/W'minP{|<5/,¿o>| > li Í I exP- 2%2 (usethe definitionof 7Y'). (2.42) Ow T O License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use ¿-DIMENSIONALRANDOM WALK 81 If 2 < T/N' < 2wK¡(d + 1) + 4, we use (2.40)instead of (2.41)and obtain, as in (2.42), P{(ST,u)>2-M( > p{(Su+xw - SJN,,ti>)>\M(u)Ak, X 0 P{\(ST,co)\>^M(U)Ak} > P{|<^,co>|> M( Jn = Jkn « /[!<*>,*>!> m(k)Ak or |,*>|>4k], vr = vk = infj«: 2 (1 - J,) - iT(*) }, yr - Y? - 2 *»• (2-46) r,_, <«<»-, Yr is the sum over a bunch of ^„'s, exactly T(k) of which satisfy \(X,ax)\ < m(k)Ak and |<.V,u2>| < .4*. It is easy to see from this that the Y„ r > 1, are independent and identically distributed. Moreover, we can write Yr= 2 XJn+ 2 xn(\-jn), License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 82 HARRY RESTEN and by definition the second term contains exactly T(k) summands with 1 - J„ ¥=0. Thus if ax,a2,..., ßx,ß2,... are independent, each a¡ (/?,) with the conditional distribution of A",,given/, = 1 (respectively 7, = 0), then 2 Xn(\-Jn) (2.47) pr_i < n < r, has the same distribution as T'k) 2 ßn- (2.48) «=i Moreover, the sum (2.47) is independent of 2, <„ = ^(k)-i + iyp{JiSsl))l(l_p[Ji=:l])r^ I > 0. (2.50) Lemma 3. There exist constants TX3(F)- TXS(F)< oo such that for 2k/i < Ak < 2*/2 and any choice of the unit vector ¿o0= ¿OqE 9C*and k > 1 one has P{\Sn\< Ak for some 2k < n < 2k+x} < Tx3{Ak2-k/2}"-2 2 p 2 1ÍV 2k-l < lx(P{Jx = 0}yXP{\Xx\>(8lxyxM(u0,k)Ak} < i,/8. A similar estimate gives for k > k7, P 2 (ßn$l) >4x0 <8' T-l, + l and we therefore conclude from (3.92) that 2 2,(ßn,Qx)-T(k)p(k) <8*J>-r (3.93) n-l However, with a¡, ßj and A independent, Y¡has the distribution of 2j_,/8fl + 2^xa( (see (2.46H2.50));(3.91) and (3.93)together therefore give P{\(Yx,ü2y\>i-6©(k)M(Uo,k),\(Yx,Üx)\< 12x0 + 3} >P{A = /0}P 2 (te) >±e(k)M(u0,k), n-l 2 Ä,ß, - P(A:)p(A;) i-i2 2 <«,#,>+ r(*)p(*)< 4x0 + 3 i-i > ^{A = 4>) • \ [ \ -/oP{||> ^2O0(A:)A/(Wo,A:)}} >¿r/P{A = /0} (see (3.70)and (3.73)). (3.94) License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use ¿/-DIMENSIONALRANDOM WALR 111 Moreover, by (2.50) and (3.34) for any /, P{A = /} > (l\)-l(TP{Jx = 1})'(1 - 2K0T~X)T > (/!)-,(2-5i,)' exp - 4/£o, as soon as K0T(k)~x < {-, i.e., for k > kn(F) if k7(F) is chosen large enough (see (2.29)).Thus, in terms of Yx,(3.94) yields for k > k-, and some K^ > 0, independent of F and k, p[vx = T+l0,\{Yx,ü2)\>i-6, |(y„ß,)|< 16(12;c0+ 3)|(y„ß2)|} > K23. We can therefore find a number xx = xx(F,k), \xx\ < 16(12x0 + 3) such that p{vx = t + /„, (y„ß2) > £, (y„ß,> < xx(y„ß2)} >\k23, (3.95) as well as p{r, = r+/0,(y„ß2)>1L, ( y„ß, ) > x, ( F„ß2)} > $ j:,,, (3.96) or both of these inequalities hold with ^{K«/0+..ß2)|<^20,|<«/0+l,ßi>|>h<«/„+l.ß2>|+i} > p{|<5/o+1,ß2>|< rnini^KnQWM^k)},^,^)^!} > P{|<5/o+„ß2>|< min{\,K20e(k)M(Uo,k)},\älo+x\> l} > 1 — e2 — 2c2 > \. Let us assume that ^{K«/0+1,ß2>|<#2» <«/0+lA> >h(«/0+A>| + 4~>Xl{%+l&l)+\}>\- (If this fails then we have p{\{%+i$2)\< *20. <«/„+!,«.) < -h<ä,0+„ß2>| -ï<^i<«/0+i,«2)-ï} >4-; License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 112 HARRY RESTEN this can be treated analogously.) Then we have from (2.50), (3.96) and (3.34), p{(y^2) >t6-k20>3l2,(yx$x) > *,(F„ß2) +{-} >P{A = l0+l] ■pU 2^+2 Ä,o2)>i, KixJ 2 * + 2 A.Qj), \ l |(«/0+i>Û2}|< K20, <«/o+i»ñi) > xx(alo+x$2) +\ , P(A = L+l} , /- \ >* P{A=l0} PiVx = T+l°' ^7,'°2) **• (y„ß,)>*,(y„ß2)} >±6K23T(k)P{Jx - l}(/0 + I)"1 > 2-^23(¿o + I)"' = J^ (3.97) say, with AT24> 0 independent of F and k. We can now complete the proof by a final application of Corollary 2 as before. (3.95) and (3.97) again are versions of (3.8) (resp. (3.9)) with Z,(l) = 2s 2^ 2^„< 12 < K.25' s> 1. (3.98) r—s i As in (3.81)-(3.83)the left-hand side of (3.98)is, however,an upper bound for the left-handside of (3.1).Thus (3.1)holds also in this last case. □ Lemmas 4 and 5 prove (3.1) with ks = max(fc6,fc7),K2 = max(Kx4,Kx9). As observed in the beginning of this section, this proves Proposition 1 and our theorem. References 1. K. L. Chung and W. H. J. Fuchs, On the distribution of values of sums of random variables, Mem.Amer. Math. Soc.No. 6 (1951).MR 12,722. 2. A. Dvoretzky and P. Erdos, Someproblems on random walk in space, Proc. Second Berkeley Sympos. Math. Statist, and Prob. (Berkeley,Calif., 1950),Univ. of California Press, Berkeley, 1951,pp. 353-367.MR 13,852. 3. H. G. Eggleston,Convexity, Cambridge Univ. Press,Cambridge, 1969. License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use ¿/-DIMENSIONALRANDOM WALK 113 4. K. B. Erickson, Recurrence sets of normed random walk in tLä, Ann. Probability 4 (1976), 802-828. 5. C. G. Esseen, On the Kolmogorov-Rogozin inequality for the concentration function, Z. Wahrscheinlichkeitstheorieund Verw.Gebiete 5 (1966),210-216. MR 34 #5128. 6. _, On the concentration function of a sum of independent random variables, Z. Wahrscheinlichkeitstheorieund Verw.Gebiete 9 (1968),290-308. MR 37 #6974. 7. W. Hengartner and R. Theodorescu, Concentrationfunctions, Academic Press, New York, 1973.MR 48 #9781. 8. K. Itô and H. P. McKean, Jr., Diffusionprocesses and their sample paths, Springer-Verlag, Berlinand New York; AcademicPress, New York, 1965.MR 33 #8031. 9. H. Kesten, The limit points of a normalized random walk, Ann. Math. Statist. 41 (1970), 1173-1205.MR 42 #1222. 10. J. V. Uspensky, Introductionto mathematicalprobability, McGraw-Hill, New York, 1937. Department of Mathematics, Cornell University, Ithaca, New York 14853 License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use|> Ak) > (8¿/+ 8)"1}. As we shall see in the next lemma t(oi) is bounded on |co|= 1, and we can therefore pick an o>d= ud which maximizes t(-,k), i.e. for which /(«£)- max t(a>,k). M"1 After that we can successively pick coJL,, ...,«,* such that M(œ,k)Ak) < K0 (2.27) and7 T(k)o2{(Xx,w)l[\(Xx,a)\< M(a,k)Ak]} < K0{M(u,k)Ak}2. (2.28) T(k)^> co and Ak inf M(a,k)-+ oo ask-*ao, (2.29) M=i ö/k/ /Aere evc/ífóa universal 0 < rj < 1 (see (2.44)) Jt/cA /Aa/ /or all \u\ = I, k> 1, P{|<5r,ío>|>¿M(co,A:)^} > r,. (2.30) Ak] < P{|S,|> 2*/8}-»0, k-*ao. Thus also for each fixed L', Ak infu M(u,k) >\L' eventually, and (2.29) holds. We turn to (2.27) and (2.28).We have for each ¿o,,3x-2M(>| >2-M(a)Ak) > (Sd + 8)"1}. Then, by (2.22), N'(u,k) < #(«,£) < T(k). Also, by the very definition of N', P{(SN.,u>-)>\M(U)Ak] > (l6)-\d+ I)"1 (2.40) or the inequality holds with (SN.,u} replaced by —(SN.,cS). For the sake of definiteness assume (2.40). Then any interval containing {SN.,cd) with a probability larger than (16¿/+ 15)(16¿/+ 16)-1 must contain points to the right of ^ M(u)Ak. By (2.39) there exists such an interval of length \ M(u)Ak which contains with a probability at least 1 - 2WK¡N'T~X> (I6d + 15)/ (16¿/+ 16) whenever T/N' > 2x*Kf(d + 1) + 4. In this case, therefore, P { \M(v)Ak < <5^,¿ú» > 1 - 2X4K2N'T-X (2.41) and P{(ST,u/>2-M(w)Ak)